Here are a few keypoints to help you navigate the prototype code:

The observation simulates a set of 40 laser proximity sensors (geometrically, they are treated as segments starting from the center of the robot which collide with the environment generating intersection points which are then given to the agent) and has the shape of an array of tuples each containing 3 informations:

• Angle: the angle of the direction the sensor was facing, relative to the environment ([0,359]);
• Distance: the distance between the agent and the eventual point of intersection found (a high number is set instead if nothing was hit);
• isObject: boolean flag telling the agent whether the eventual point hit an objective or not (assumes the agent can recognize objectives);

There are 3 possible actions: rotate right by 45°, rotate left by 45° or move in the faced direction.

The task is to collect as many objectives as possible within 1500 time steps (approximately 15 seconds). The agent provided in the code succeedes in completing the task and definitely outperforms a human player. Reward is 1 for every objective collected.

The first expansion of this work would be to complete the above mentioned task moving the objectives spawn points inside the environment rooms. This already seems to require a formulation for curiosity as seen in a recent paper published by DeepMind. (https://github.com/google-research/episodic-curiosity) Next, once the agent is comfortable exploring a single environment while finding objectives, one could attempt to train it in a set of 20-30 environments (switching it with each new episode) until it's capable of performing the same task on a new one which it has never seen before.

Please make sure these dependencies are installed first:

• Keras
• Pyswip (so the latest version of SWI-Prolog as well)
• Pygame
• Guizero
• Numpy

Then clone our repo and simply run Main.py.

For a quick glance at what an environment looks like Prolog wise, JSON wise and graphically, please see the example1 in the environments folder.